给 Claude Code / Codex 加 Jev 工具调用护栏

st1ne @SolSt1ne原推发布于 2026.09.27

在 X 上打开原推

this is pure gold for anyone running Claude Code or Codex Jev founder Diogo Amogo wrote a note on putting a Jev guardrail layer around every tool call your agent makes the problem: > your agent can run rm -rf, read your .env and force-push to main > most setups either approve everything or ask you about everything > or worse, ask the same model to judge its own command the fix is one typed call before every command: > destructive? > touches secrets? > outside the task? > how big is the blast radius? > allow, ask or block? 5 answers, 1 request, under a second the model never gets the last word. a decision table in code does - what it catches > cat .env → blocked > DROP TABLE users → blocked > rm -rf ./dist → sent to you with the reason > edit src/router.ts → runs, then checked against the diff - how to use it > drop this note into Claude Code or Codex > let it wire the guard into its own hooks one evening of setup and your agent stops asking you about ls and stops touching things it shouldn't ↓

这条是引用着另一条写的：x.com/solst1ne/status/2101724066369654935